A new generative approach for optical coherence tomography data scarcity: unpaired mutual conversion between scanning...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) presents a trade-off between scanning speed and image quality.
• Different OCT scanning presets yield visually distinct images, hindering data compatibility and analysis.
• High-quality OCT data is scarce, impacting the development of robust diagnostic systems.

Purpose of the Study:

• To develop an automatic methodology for the visual conversion between prevalent OCT scanning presets.
• To enable the transformation of low-quality OCT images into a high-visibility style and vice-versa.
• To enhance the compatibility of OCT datasets for training artificial intelligence models.

Main Methods:

• Utilized contrastive unpaired translation generative adversarial architectures for image conversion.
• Developed two generative models for mutual conversion between Macular Cube and Seven Lines presets.
• Preserved natural tissue structure while modifying the visual appearance of OCT images.

Main Results:

• Synthetic generated OCT images achieved quality scores comparable to original target preset images (BRISQUE).
• Generative models successfully replicated the visual characteristics of original images in separability tests.
• The methodology demonstrated effective unpaired visual conversion between OCT scanning presets.

Conclusions:

• The proposed methodology can create multi-preset OCT datasets for training more robust computer-aided diagnosis systems.
• This approach addresses the scarcity of high-quality OCT data by leveraging existing datasets.
• Enables AI systems to be trained on diverse visual features encountered in real clinical settings.